Our minds aren’t equipped to handle AI

Writing in The Verge, education writer Benjamin Riley argues that the AI industry is working from a bad model of the human mind. The brain-as-computer metaphor — input, computation, output — drives how labs build and sell AI. Riley leans on neuroscientist Paul Cisek’s view that brains are better understood as feedback-control systems shaped by evolution and, crucially, by social learning and institutions like schools.

His conclusion is blunt: AI is tempting in the moment and harmful as a steady diet, especially for learners. He points to OpenAI’s claim that learners make up more than half of ChatGPT’s users, Anthropic’s “Claude for Teachers” initiative, evidence that students stop doing homework once they start using AI, and new bans — Norway for students under 13, plus districts such as Los Angeles and New York City.

This matters because the fight over AI in classrooms is moving from opinion pages to policy. Bans are spreading, and the industry is racing in the opposite direction.

We don’t buy the whole thesis. Calling AI junk food for the mind flattens a tool that, used well, can explain a hard concept a dozen ways, give a stuck student a next step, or help a teacher who has 150 kids and no time. Plenty of people use these systems to learn more, not less. A blanket diagnosis that the technology itself is the poison isn’t supported by the evidence any more than the hype is.

But Riley is right about the thing that matters most. Learning requires effort, and people take off-ramps when they’re offered. The essay cites researchers describing cognitive delegation: as people hand off reasoning, the social environment that supports independent thinking weakens, which makes handing off reasoning even easier. That is a credible feedback loop, and it is exactly what a frictionless homework machine invites. A tool that does the work isn’t a tutor. It’s a substitute.

I’m also struck by how the industry frames this. Counting students as the biggest slice of ChatGPT’s users is presented as a triumph. It could just as easily be a warning sign. If a big share of that usage is students skipping the hard part, the metric is measuring erosion, not education.

Where we land: the answer isn’t to pretend AI doesn’t exist in kids’ lives, and it isn’t to flood classrooms with free premium chatbots either. The useful version of AI in schools is designed around friction — it asks questions back, refuses to just hand over the essay, and gives teachers visibility. Labs that want credibility in education should build that, measure learning outcomes rather than engagement, and publish the results. If they keep optimizing for usage, the bans Riley welcomes will keep spreading, and frankly they’ll have earned them.

Riley discloses that he has informally advised Schools Beyond Screens, a group that pushed for school chatbot bans. The essay also cites François Chollet’s description of AI as cognitive automation and a scholarly argument that LLMs are better seen as cultural and social technologies than as intelligent agents.